Decide what an AI agent is allowed to do

Separate reading, drafting and acting

An agent may read information, prepare a proposal or change a live system. Those are different permissions. Begin with the narrowest capability that solves the task. An assistant preparing a report does not need permission to delete customer records or change billing.

Write an action list

For each proposed action, identify the system, required data, allowed output, reviewer and exception route. Specify what is prohibited. For example, a customer-support assistant may draft a reply from approved sources while refunds remain outside its access.

Limit the data as well as the action

Choose only the records needed for the task. Confirm where information is processed, who can see it and how access can be withdrawn. Do not copy an entire mailbox simply because it is easier than selecting a relevant folder or collection.

Test refusal and recovery

Try a request outside the agreed scope, a missing record and a disconnected account. The agent should stop or ask for review rather than improvise. Check whether the integration supports restricted credentials and whether actions can be logged without exposing sensitive content.

The Privacy Commissioner provides guidance on AI and personal information. Privacy Commissioner of Canada: AI and privacy

Price the operating responsibility

Include permission reviews, credential maintenance, incident handling and staff training in the project scope. Measure completed useful work and corrections, not just the number of actions. A smaller assistant with dependable boundaries may be worth more than a broad agent that nobody is comfortable using.

Discuss an implementation

This guide applies across Canada. We provide remote AI consulting for businesses in Toronto; it does not describe a local client or a staffed office.

AI consulting in Toronto · Explore the service · Compare setup options